Having a busy summer, as always, but here’s a brief note: typos are a problem for machine translators. The best ones (like DeepL) have some ability to recognize typos and correct for them. But they usually can’t do that when the typo is a correctly spelled word (just not the one you intended to type).
Case in point: DeepL rendered a German sentence into English as “The magazine is available oh offline.” When I checked the source text, it turned out the author had mistyped “auch” (also) as “ach” (oh). The magazine was, in fact, “also available offline.”
I find these things rather delightful, especially because they score a point for Team Human…for now.
Machine translation systems still have trouble recognizing proper names. Recently I was proofreading English texts for an Austrian museum and came across a line that went something like this:
The project was directed by Thomas Schmidt and Elderberry Hot.
Now, a lot of the texts I work with are about artists and some of them have unusual names like Friedensreich Hundertwasser or VALIE EXPORT, so on the first read-through I thought, “Elderberry Hot? OK, sure, whatever!”
Then it hit me: whoever submitted this to the museum may have written it in German, used a machine translator like DeepL or Google Translate, and passed it along without checking for errors.
I translated the name “Elderberry Hot” into German and searched the Internet for it. And indeed, there was a person with that name at the relevant institution. So I corrected the English document accordingly.
I won’t tell you Elderberry Hot’s real name here, because I don’t want to affect his Internet search results. But you can figure it out with a dictionary.
I like to test DeepL’s handling of names. Analyzing passages from a book I worked on this summer, it decided a man with the surname Kicherer should be called “Giggles” (“Local politicians attended the opening ceremony and Giggles was honored”) and described a scientific report on plastics written by one Mr. Keim as “the germination report.”
Yet more evidence that machine translation must always be supervised by humans with real expertise. If you wouldn’t leave a two-year-old alone in your office, you shouldn’t trust MT to handle your documents on its own, either.
Today I ran across an example of a category most people don’t know about: extremely close translations for opera singers.
I say “close” rather than “literal” because it’s not just about communicating the exact meaning, but also keeping words in mostly the same order so that you could basically nail the target sentence on top of the source sentence and they’d match up nicely. Here’s one from my own sheet music collection:
By conventional standards, this is the worst translation ever. But it’s just what singers need if they aren’t fluent in the language of the song (which many of them aren’t, especially if they’re just starting out). They have to know what each word means so they can sing them with appropriate emphasis and expression.
I unexpectedly ran across one of these today. I was searching for the libretto of The Merry Widow online to see how previous translators had handled a pun in the spoken dialogue. The first result was this.
Opening lines:
Verehrteste Damen und Herren | Most honorable ladies and gentlemen Ich halt es für Gastespflicht, | I hold it as duty of a guest Den Hausherrn dankend zu feiern, | The host of the house thankfully to fete Doch Redner – das bin ich nicht! | Yet, a speaker – that I am not!
At first I thought it was just a terrible translation, because most of the libretto translations available online are intended for listeners and readers, rather than singers – in other words, they’re mostly “good” translations for people who want to read along with a recording or study the libretto as literature. I wondered if it was an MT (e.g. Google Translate) job but that didn’t seem quite right. The translator’s name appeared at the end and I looked her up – it turns out she was a language coach at Opera San Jose, which explains everything. She was not (as I initially suspected) a crazy person, but rather someone with the very specific job of helping opera singers understand their lines with maximum accuracy.
The reason I thought it probably wasn’t MT is that machine translators never follow source-language word order that closely. So you actually can’t get an MT to do a job like that. It’s programmed to translate into normal English word order, not the very close word order needed for these learning aids.
Here’s how 1. Google Translate and 2. DeepL render the opening lines:
1. Dearest ladies and gentlemen, I consider it a guest obligation To thank the host, But speakers – I am not!
2. Dear Sirs, I consider it a guest duty, To celebrate the landlord with thanks, But I’m not a talker!
I like how DeepL assumed it was a letter.
Anyway, both of those are terrible. But – happily for language coaches at the opera – neither of them is terrible in the right way.
I often run legal boilerplate (with all specifics of the case removed) through DeepL as a preliminary measure. Today’s experiment was from an alimony contract. And I’m pretty sure the person who signed it didn’t intend to agree to this:
Because of the fulfilment of the obligations arising from this document, I submit myself to immediate execution.
(Wegen der Erfüllung der Verbindlichkeiten aus dieser Urkunde unterwerfe ich mich der sofortigen Zwangsvollstreckung.)
Some people seem to think (see this post) that a really great machine-translation program would be able to “handle complicated multilingual puns with ease.” But what is a “multilingual pun” anyway?
The prefix “multi” implies more than two, and honestly, off the top of my head I can’t think of any puns involving more than two languages. If you have one, please send it in!
But I know of some puns between two languages – I think these are properly called “interlingual puns.”
For example, here’s an old joke my Dad used to enjoy telling: “What did the alien say in the music store?” – “Take me to your Lieder.”
Ha ha ha ha ha ha ….sigh.
Consider whether an AI machine translator could handle this pun with ease – by “handle” I assume the author meant “translate.” Could the best ever future MT translate this joke into Japanese, or Thai, or French, or Navajo, or Spanish? Um, no? Neither could a human translator. It’s just not possible. If it appeared in a story you were translating, you’d insert a similar joke, but you couldn’t reproduce this exact joke. Some things really are not translatable.
The joke works by combining English and German, so could you translate it into German? Again, no. It wouldn’t work in reverse, so to speak. German speakers with good English and a good knowledge of 20th century pop culture would be able to get it, though.
OK, let’s try another one. I saw this pun being assembled in real time on social media. A friend of mine who was getting a PhD in Arab Studies posted:
ABD
which his academic pals recognized as the acronym for “all but dissertation,” i.e. he had reached an important phase in his course of study. Congratulations were offered in the comments section, but someone also left this comment:
abd al-dissertation
which is a pretty funny pun because it sounds like an Arabic name meaning “servant of the dissertation,” and that of course is what he was going to be for the next year or so. (For more context on the Arabic name in question, go here.)
Is this translatable into any other language? Nope. And there’s nothing humans or MT can do about it. Part of the romance of translation is the bittersweet knowledge that some things just can’t be carried over into another language, like rare flowers that won’t grow outside their native land.
In this post I promised to go through some pun-translation strategies.
What makes puns hard to
translate is that there is almost never one “right” or “best” solution. Puns
give rise to several different scenarios:
1. You just translate the straight meaning and write a footnote about how it was a pun in the source text. Sad, right? But very common in certain contexts, e.g. academia. I actually had to do it last week.
On the other hand, some
academic translators do get creative, especially if the goal of the translation
is twofold: to inform readers and to give them an experience of the text that
parallels the original. Erika Rummel does this in her translation of
Reformation dialogues, e.g.:
LEGATE: I also confer doctorates.
BRUNO: Donkey doctorates.
That line has a footnote, which reads: Literally, “troubles, not doctorates”; the pun dolores/doctores cannot be rendered into English.
For a dialogue that you might want to read out in class and have some fun with, inserting a new joke is a good idea. But she still has to explain the original joke in the footnote so students can be fully informed about the content of the source text.
2. If you’re working with
related languages, you might get lucky: for example, Kurt Schuschnigg’s rhyming
declaration “Bis in den Tod! Rot-Weiß-Rot!” is easily translated as
“Red-White-Red until we’re dead!” because historical linguistics has done your
work for you.
3. You can think of a similar pun using different words. In the 2015 film Er ist wieder da, main character Sawatzki thinks a rat is pregnant; the rat is actually male and its owner says: “Die sind die Eier.” (Lit. “Those are the eggs” with “eggs” being German slang for testicles.) Sawatzki, still confused about whether the rat is male or female, responds, “Die Ratten legen Eier?” (“Rats lay eggs?”). In English, your choices are “nuts” or “balls,” so the confusion over laying eggs is out. Instead, the subtitler came up with “Those are his nuts” – “Rats collect nuts too?” which is pretty good. (You can watch this movie on Netflix, by the way, as Look Who’s Back. It’s actually more about Hitler than it is about pet rats.)
4. In some cases, you have
room to think of something very different from the original. In Fontane’s Effi
Briest, Effi’s cousin tells a lame joke about Job because Bible jokes are
all the rage in Berlin:
»Die Fragestellung – alle diese Witze
treten nämlich in Frageform auf – ist übrigens in vorliegendem Falle von großer
Simplizität und lautet: ‘Wer war der erste Kutscher?’ Und nun rate.«
»Sehr gut. Du bist doch
ein Daus, Effi. Ich wäre nicht darauf gekommen. Aber trotzdem, du triffst damit
nicht ins Schwarze. «
»Nun, wer war es denn?«
»Der erste Kutscher war
‘Leid’. Denn schon im Buche Hiob heißt es: ‘Leid soll mir nicht widerfahren’,
oder auch ‘wieder fahren’ in zwei Wörtern und mit einem e.«
OK. Basically, cousin Briest asks “Who was the first coachman?” and the answer is “sorrow,” because in the Book of Job it says “Sorrow shall not befall me” and in German the word for “befall” is “widerfahren,” which sounds just like “wieder fahren,” which in turn means “to drive again.” So, “Sorrow shall not befall me” and “Sorrow shall not drive me again”* sound alike in German, hence the joke.
As you can see, the original pun is completely untranslatable. What to do?
Here’s how Helen Chambers
and Hugh Rorrison handled it in their translation for Penguin Classics:
‘The question in this case
— all these jokes take the form of questions by the way — is of the utmost
simplicity: “What was our Lord’s favourite plaything called?” Now guess.’
‘Little lambkin, perhaps.’
‘A brave try. You’re an
ace, Effi. I’d never have thought of that. But you’re wide of the mark.’
‘Well, what was it then?’
‘Our Lord’s favourite plaything
was called “Gladly”, because in the hymn it says ‘Gladly the cross I’d bear” or
“cross-eyed bear”, “eyed”, e-y-e-d.’
They had to find a completely different joke that was equally cringey and also related to the Bible. Other translators would have thought of something else again — many of these would work (I say “many” because you have to check that the joke would have made sense in 1895, when Effi Briest came out).
When I first read this, I agreed that Chambers and Rorrison’s joke had the right level of lameness, but thought it erred in not actually being a “Bible joke” per se. Cousin Briest explicitly introduces it as a Bible joke and says the pun comes from the Book of Job, but “Gladly the cross I’d bear” is from a hymn. However, after much searching — searching through German websites that offer the full text of Luther’s Bible, but also global Google searches with and without quotation marks and with variations in phrasing — I don’t think cousin Briest’s punchline is actually a Bible verse at all. Apart from Effi Briest, the only place I found it was in this forum, where it’s attributed (falsely, I think) to the book of Daniel:
Wer war der erste Berlina? Das war Daniel in der Löwengrube: “Leid soll mir nicht widerfahren.” Damit ist auch die Frage beantwortet wer der erste Kutscher war -> Leid
Similarly, “Gladly the cross I’d bear” seems to be a misquotation from “Keep Thou My Way” by Fanny Crosby. So all in all, this joke matches both the style and the dubious sourcing of the original joke quite perfectly.
Those are a few examples of pun translation. Now, apropos my earlier post about how good MT could get, do you think an MT could ever deal with this problem? I don’t.
*********
*Although, does “Leid soll mir nicht wieder fahren,” actually make sense with that dative “mir”? Is it just an imperfection that makes the joke extra lame? Or is the speaker being driven into? Usually when you drive someone somewhere, that someone is in the accusative.
Here’s the start of a German sentence I’m working on right now:
Die überwältigende Musik in Kombination mit kurzen, pathetisch vorgetragenen Deklarationen von hehren Zielen …
And here’s how the best free machine translator renders it into English:
The overwhelming music combined with short, pathetic declarations of noble goals …
But according to dict.cc, “pathetisch” could be translated as solemn, emotive, histrionic, pathetic, lofty, dramatic, impassioned, melodramatic, emotional, or declamatory. Sounds like we need an actual human to reflect on the context here and make an informed choice.
Over at Slate Star Codex, Scott Alexander has a good post about the future of AI, but I need to nitpick these speculations about what a “future superintelligent Google Translate” could do:
For example, take Google
Translate. A future superintelligent Google Translate would be able to
translate texts faster and better than any human translator, capturing
subtleties of language beyond what even a native speaker could pick up. It
might be able to understand hundreds of languages, handle complicated
multilingual puns with ease, do all sorts of amazing things.
This description raises interesting
questions about what the best possible machine translation (MT) would be like.
Let’s go through it point by point:
Is MT faster than any
human translator? Yes.
Can it “understand hundreds of languages”? No. The problem with MT is that it doesn’t actually “understand” any languages in the sense humans do. It matches patterns. For more on this topic, see Scott Spires’ post “Machine translation and savant syndrome” or my posts “Senta spinnt” and “Easy for humans, hard for computers.” (As an update to that post, I should say MT programs are getting better at dealing with common typos. But I think my basic point still stands.) What would have to happen for MT to truly understand a language? It would have to be an entity as complex as Star Trek’s Data – something that moves through the world, interacts with people, has experiences (including experiences of real-world communication and miscommunication) and personal memories – and even he has trouble sometimes.
Could the best possible
future MT “capture subtleties of language beyond what even a native speaker
could pick up?”
Seems unlikely.
Now, there are aspects of
language that machines measure more accurately than humans. For example, they
can measure the resonance of the phonemes you produce in cycles per second. An
AI can store tons of vocab, which means it can make very precise matches very
quickly. An AI with a huge data set of spoken language could analyze very
subtle aspects of speech most humans would miss. It might be able to conclude
from your speech that you’ll be diagnosed with a neurological disorder next
year, or guess your age with almost perfect accuracy. There’s probably already an
AI that does this kind of thing.
But what would it take for
an artificially intelligent translator of written texts to pick up more from a
text than a human could, or capture more of its subtleties? What would this
look like, and what inputs would be needed? What would be in your training data
set and what instructions would you give the AI?
To start with, you could feed it tons and tons of books. For German-to-English MT, you’d take every published book that exists in both English and German, and present them to your AI in pairs. For example, DeepL would not have made the error in “Senta spinnt” if its training data had included the original German libretto of The Flying Dutchman and a good English translation of same. An AI trained on all available pairs of translated literary classics would outperform human translators at identifying literary quotations, and if you gave it Schlegel’s German version of Hamlet , it would recognize that for what it is and give you Shakespeare’s Hamlet rather than this:
To be or not to be; that is the question:
Obs nobler in the mind, the arrow and spin
Endure the angry fate or,
Wielding against a sea of plagues,
By resistance they end? Dying – sleeping –
Nothing else! And to know that a sleep
The heartache and the thousand blows ends,
Our meat’s heritage, it’s a goal
To wish for the most intimate. Dying – sleeping –
Sleep! Maybe dream too! Yes, there is.
So MT could get better at recognizing existing translations. But of course, a large set of training data also helps MT to create good new translations of its own. DeepL has access to masses of web content as training data, which is why it’s so good at translating boilerplate:
For Germany’s CDU, the
following applies: we must resolutely combat climate change and implement the
Paris Agreement consistently. Strong climate protection legislation is the
foundation on which we can credibly achieve our goals. We take this seriously
and clarify how, for example, a “CO2 cap” with a binding climate
protection path in the form of a national certificate trade could be
implemented in the near future, particularly in the areas of transport and
buildings. [press release translated by DeepL.]
That’s decent, but it’s
not better than what a human would do, and could it ever be? If DeepL
analyzed all the press releases ever written, in what way would its output be
consistently better than the best human translators? I think it would be a slightly
improved version of what it is now: much faster, and almost as good.
What specific aspects of MT could get substantially better than they are now? One area with strong potential for improvement is matching styles to time periods. I can imagine a future MT where you could select a time period for the text you’re putting in, so that, say, the MT wouldn’t translate “Mama und Papa” from a nineteenth-century text as “Mom and Dad”. I can also imagine one that would convert German footnotes into MLA or APA style in English. Both of those are useful but they’re also things humans already do, so again, the MT wouldn’t be outperforming humans.
As far as I can tell, there’s a ceiling for MT improvement set by the MT’s total lack of knowledge and experience. It doesn’t know anything and it’s never done anything, been anywhere, or met anyone. It never will. It doesn’t have a theory of mind enabling it to guess whether the average reader will find a given sentence easy or hard to understand and to adjust its phrasing accordingly. It doesn’t know that certain turns of phrase might annoy certain kinds of people. It doesn’t know whether its translation of an ad grabs people’s attention or not. It doesn’t know who is feeding it a text or what they plan to do with the translation. (It could have some information about those issues – e.g., a metric that says “NSFW” is an attention-grabbing term – but it wouldn’t have the understanding required for human-style judgement calls.) What it can do is analyze lots of data about which words and phrases in different languages correspond to each other under what textual circumstances and…isn’t that it? Apart from speed and stored vocab, in what specific way could it actually exceed human translating ability? That’s an honest question, so if you have an answer, please comment.
On to the last point: Could this future MT handle complicated multilingual puns with ease? This assertion is really interesting because puns are among the hardest aspects of translation. And they offer a broad scope for action that ranges from essentially doing nothing to making really wild choices. I wish Mr. Alexander would come over here and explain how he imagines such a thing would be accomplished.
There’s a lot to say about puns and this post is already too long. So I’ll write a follow-up post about puns and jokes…I have plenty of material. It’ll be fun. Until then, please comment with your ideas about how good MT could really be.
People tend to think of machine-translation post editing as “easier” than old-fashioned translation.
But I’ve come to think of MTPE not as easier than traditional translation, but as requiring a different set of skills. Namely, the same skills required by find-the-difference puzzles.
Find-the-difference puzzles vary in difficulty, of course, just as MT jobs do. You might — especially if you’re working with an extremely cheap online translation service — get an MT job consisting of single sentence with one glaring error, in which case it feels like this puzzle:
Or you might get a longer, more complex text with subtler errors, like this puzzle:
The worst scenario is a long text where the MT has done a mostly acceptable job. You compare 10 sentences in a row and they’re all fine. This makes you lazy and then you overlook the errors that do crop up from time to time. It’s like getting two versions of a crowded Brueghel painting and the only difference between them is that one person is missing a shoe.
In any case, the mental processes required for MTPE are different from those required by traditional translation. I was going to say that translating the old-fashioned way is like unraveling a knitted garment and knitting it back up again according to a different pattern. But since my first metaphor was a find-the-difference puzzle, I should compare it to a puzzle. Maybe it’s like this puzzle:
Both jobs have the same goal, which is to produce a text that is correct and comprehensible in another language. They also require similar basic skills — in both cases you need to have a good understanding of two languages and decent writing skills in the target language (although it’s worth noting that MTPE doesn’t require excellent writing skills, just “premium mediocre” ones) but your brain is doing a different kind of job in each case. So I wouldn’t be surprised if some translators are better at one than the other. Perhaps the future of the industry will see a sharp division between find-the-difference people and rearrange-the-shapes people, handling different kinds of texts.
While cleaning out a desk drawer I found these amusing machine translation errors I’d noted down months ago. So here they are (sorry about the line breaks within words):
Mensch das ist super lieb.
Human that is super nice.
Man, that’s really nice.
Ich war nur ehrlich und habe jetzt den Salat.
I was just honest and now I have the lettuce.
I was just being honest and now I have to deal with the consequences.
Liebe Frau Knopf
Dear Mrs. Button
Dear Mrs. Knopf
Ich bin leicht sauer.
I’m slightly acidic.
I’m a little angry.
Darüber bin ich sehr erbost und weiß aber auch das [sic] Fehler passieren können.
I’m very angry about that and white but also the mistakes happen.
I’m very angry about it, but I also know that mistakes happen.
